As quantum technologies gain relevance across scientific and industrial domains, accessible educational frameworks are critical to preparing the next generation of quantum-literate learners. Traditional instruction often fails to engage diverse audiences or convey abstract quantum concepts intuitively. This paper introduces a novel AI-assisted learning architecture centered on the Embodied Language Model (ELM) -- a hybrid approach that integrates Reinforcement Learning (RL) and Large Language Models (LLMs) -- to address key challenges in quantum education through adaptive, game-based learning. We evaluate this framework in Qookies, a story-driven point-and-click adventure featuring the AI-controlled non-player character (NPC) Yuki as co-learner. Yuki combines an RL-based action model with an LLM which uniquely, like the player, begins with limited domain knowledge, acquiring understanding incrementally through shared game-play. The interaction design emphasizes observational learning, instruction, and dialogue: the player prompts Yuki to act, requests assistance, or engages in open conversation, while Yuki defers to the player when uncertain, suggests interactions with objects outside her reach, or explains quantum concepts contextually. Key contributions include the RL model’s runtime learning of object concepts through game-play observation, the LLM’s prompt accumulation by narrative progression and user dialogue, and the interface design enabling communication between RL and LLM models, collectively mirroring the player’s learning progress. Our architecture integrates these complementary processes to enable adaptive, personalized learning through collaborative exploration. Evaluation results from related studies suggest that the game enhances learners’ conceptual understanding of quantum phenomena, whereas the grounding RL component effectively reduces intrinsic cognitive load, simplifying the acquisition of complex concepts and promoting lasting learning through contextualized game-play. Overall, our framework offers a scalable, adaptive model for AI-assisted personalized education, contributes to hybrid AI architectures in educational technology, and suggests potential for domain transfer across STEM learning environments.